Dynamical Update Maps for Particle Flow with Differential Algebra
Simone Servadio · 2025
Read Dynamical Update Maps for Particle Flow with Differential Algebra by Simone Servadio (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-01598.txt.
Introduction
Simone Servadio speeds up a class of Bayesian filters used to track uncertain systems, such as a small satellite working out its orientation. Particle flow filters move a cloud of sample points so that it matches incoming measurements, but the math of moving each point is costly. Servadio's idea is to use differential algebra, a method that represents the flow of the system as polynomial maps built from deviation vectors. Two sets of maps result: one pushes the particles forward in time, and the other performs the measurement update.
With both steps reduced to evaluating polynomials, the filter runs much faster. Tests on numerical examples show the time savings, which matter for small systems such as CubeSats with weak onboard computers. The filter's structure stays intact; heavy calculations are replaced by map evaluations, trading algebra done once, up front, for speed at run time.
Catalog record
| CREATOR | Simone Servadio |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-01598
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Dynamical Update Maps for Particle Flow with Differential Algebra Dr. Simone Servadio, Assistant Professor, Department of Aerospace Engineering, [email protected] Simone Servadio Abstract Particle Flow Filters estimate the “a posteriori” probability density function (PDF) by moving an ensemble of particles according to the likelihood. Particles are propagated under the system dynamics until a measurement becomes available when each particle undergoes an additional stochastic differential equation in a pseudo-time that updates the distribution following a homotopy transformation. This flow of particles can be represented as a recursive update step of the filter. In this work, we leverage the Differential Algebra (DA) representation of the solution flow of dynamics to improve the computational burden of particle flow filters. Thanks to this approximation, both the prediction and the update differential equations are solved in the DA framework, creating two sets of polynomial maps: the first propagates particles forward in time while the second updates particles, achieving the flow. The final result is a new particle flow filter that rapidly propagates and updates PDFs using mathematics based on deviation vectors. Numerical applications show the benefits of the proposed technique, especially in reducing computational time, so that small systems such as CubeSats can run the filter for attitude determination. I Introduction The estimation problem consists of optimally merging information from noisy observations with prior information, in terms of a distribution, to obtain a representation of the posterior probability density function (PDF). When the propagation of PDFs is included, we obtain a full filtering algorithm. In the linear and Gaussian case, the optimal solution is the Kalman filter [ 1 ] . However, most systems are nonlinear and require different techniques to obtain an approximation of the distributions: classic examples are the Extended Kalman Filter (EKF) [ 2 ] and the Unscented Kalman Filter [ 3 ] . Particles represent an accurate representation of the PDF, where their positions and weight approximate a continuous distribution as a probability mass distribution (PMF). When a measurement becomes available, particle filters modify the weights of the particles to match the posterior PDF, such as the Bootstrap Particle Filter (BPF) [ 4 ] and the Gaussian Particle Filters (GPF) [ 5 ] . Sequential Importance Sampling (SIS) tries to populate the posterior region with particles by introducing an importance distribution to sample from and successively modify their weight according to the likelihood [ 6 ] . Another approach of particle filters is to migrate particles from their initial (prior) location to the posterior, constituting particle flow filters [ 7 , 8 ] , and more recently [ 9 ] . Particle flow filters use a homotopy transformation that modifies the prior to the posterior gradually. This gradual change can be interpreted as a recursive measurement update [ 10 ] , which can be applied to every single particle of the ensemble [ 11 ] . This paper proposes a new type of particle flow based on the recursive update formulation derived in the Differential Algebra (DA) framework to gain computational advantages without decreasing accuracy. The DA approximates functions with their Taylor expansion series, and it is embedded with numerical integration and analytical differentiation [ 12 ] . Its use has been proven beneficial for space applications [ 13 ] , especially for uncertainty propagation [ 14 ] and quantification [ 15 ] . Indeed, DA has provided an approximation of the solution flow of a dynamical system, introduced as the state transition polynomial map (STPM) [ 16 ] . The STPM has been exploited to propagate (and update) central moments [ 17 , 18 ] . The new particle flow derived in this work exploits the DA representation of variables and the composition of the functions to propagate and update an ensemble of particles with one single mathematical operation: a computationally fast polynomial evaluation. The result is a fast particle flow filter that is not weighted down by the expansive numerical integration of particles. The resulting algorithm is applied to the attitude determination problem [ 19 ] for small spacecraft, where computational light filtering software is a necessity. The paper is organized as follows. A general explanation of particle flow filtering is offered in Section II , with the derivation of the flow differential equations via recursive update derived in Section III . Section IV presents the capabilities of DA, which are then applied to the recursive update in Section V . Section VI shows how DA can also be employed in the prediction step of the filtering algorithm, with the possibility of combining polynomial maps, Section VII . Section VIII shows the application to a toy problem, while Section IX applies the newly derived filter to the attitude determination problem. Lastly, conclusions are drawn in Section X . II The Particle Flow Update Consider some prior knowledge of the state of a system, described according to the prior distribution, p_{\mathbf{x}}(\mathbf{x}) , and a known nonlinear measurement model in the form of \mathbf{y}=\mathbf{h}(\mathbf{x})+\mathbf{v} (1) with Jacobian \mathbf{H}=\dfrac{d\mathbf{h}(\mathbf{x})}{d\mathbf{x}} (2) where \mathbf{y} is the m -dimensional measurement, \mathbf{x} is the n -dimensional state, and \mathbf{v} is zero-mean additive noise, assumed Gaussian with known covariance matrix \mathbf{R} . Whenever a measurement outcome is provided, Bayes’ rule provides the posterior probability density function, which merges measurement information with the prior PDF according to p_{\mathbf{x}}(\mathbf{x}|\mathbf{y})=\dfrac{p_{\mathbf{x}}(\mathbf{x})p_{\mathbf{y}}(\mathbf{y}|\mathbf{x})}{p_{\mathbf{y}}(\mathbf{y})} (3) where p_{\mathbf{y}}(\mathbf{y}|\mathbf{x}) is the conditional distribution of the measurement given the state, called likelihood , and p_{\mathbf{y}}(\mathbf{y}) is the normalizing constant, the marginal distribution of the measurement, such that the posterior integrates to unity in its domain. After taking the log function, the Bayes’s formulation becomes \log p_{\mathbf{x}}(\mathbf{x}|\mathbf{y})=\log p_{\mathbf{x}}(\mathbf{x})+\log p_{\mathbf{y}}(\mathbf{y}|\mathbf{x})-\log p_{\mathbf{y}}(\mathbf{y}) (4) The particle flow measurement update provides the posterior distribution as an ensemble of particles that move iteratively. Starting from Eq. ( 4 ), particle flow defines a scalar homotopy parameter 0\leq\lambda\leq 1 and the relative log-homotopy in the form of \log p_{\mathbf{x}}(\mathbf{x}|\mathbf{y},\lambda)=\log p_{\mathbf{x}}(\mathbf{x})+\lambda\log p_{\mathbf{y}}(\mathbf{y}|\mathbf{x})-\log K(\lambda) (5) where the normalization constant K has been parametrized by \lambda to ensure that the posterior distribution is a valid PDF that integrates to unity. It is evident that p_{\mathbf{x}}(\mathbf{x}|\mathbf{y},0) reduces to the prior distribution while p_{\mathbf{x}}(\mathbf{x}|\mathbf{y},1) is the posterior PDF, as it represents the classic Baye’s formulation. Particles move starting from the null pseudo-time condition until \lambda=1 , where the prior has been transformed into the posterior. Thus, this update takes place in this pseudo-time domain, following a stochastic differential equation that defines the flow of the particles: d\mathbf{x}=\mathbf{f}(\mathbf{x},\lambda)d\lambda+\mathbf{B}(\mathbf{x},\lambda)d\mathbf{w}_{\lambda} (6) where \mathbf{f}(\mathbf{x},\lambda) is the drift function, \mathbf{B}(\mathbf{x},\lambda) is the diffusion matrix function, and d\mathbf{w}_{\lambda} is a Weiner process that accounts for the stochastic nature of the differential equation. When \mathbf{B}(\mathbf{x},\lambda) is null, then the flow becomes deterministic, where each particle is governed by the following differential equation \dot{\mathbf{x}}(\lambda)=\dfrac{d\mathbf{x}}{d\lambda}=\mathbf{f}(\mathbf{x},\lambda) (7) Different approaches and techniques derived various formulations of the flow \mathbf{f}(\mathbf{x},\lambda) , such as the Gromov flow [ 8 ] and the exact flow [ 7 ] . III Particle Flowing with Recursive Update The implementation of particle flow via the homotopy transformation of the prior distribution into the posterior can be interpreted, from a different perspective, as a recursive measurement update with inflated noise. Following the information filter formulation, the information state is defined as \displaystyle\mathbf{S} \displaystyle=\mathbf{P}^{-1} (8) \displaystyle\mathbf{z} \displaystyle=\mathbf{S}\mathbf{x} (9) where